void-2024/PosterCopilot
PosterCopilot-7B, developed by Jiazhe Wei et al., is a 7 billion parameter model based on Qwen/Qwen2.5-VL-7B-Instruct, designed for graphic design layout reasoning. It takes 2-25 RGB layer assets and a canvas size as input, outputting a JSON object with bounding box and stacking order for each layer. This model excels at generating professional poster layouts, offering a unique capability for automated visual design.
Loading preview...
Overview
PosterCopilot-7B is a 7 billion parameter model, developed by Jiazhe Wei et al., specifically engineered for graphic design layout reasoning. Built upon the Qwen/Qwen2.5-VL-7B-Instruct base, it processes multiple RGB layer assets and a target canvas size to generate a complete poster layout in JSON format, including bounding box coordinates, dimensions, and stacking order for each layer. This model was accepted to ECCV 2026, highlighting its novel approach to automated visual design.
Key Capabilities
- Automated Layout Generation: Transforms a collection of image assets and a canvas size into a structured poster layout.
- Detailed Output: Provides precise
x, y, w, h, order, categoryfor each layer, enabling programmatic rendering. - Visual-Reality Alignment: Trained using a multi-stage process including perturbed SFT and RL from aesthetic feedback to ensure high-quality, visually appealing outputs.
- Controllable Design: Supports optional design briefs to guide the layout generation process.
Good for
- Graphic Design Automation: Ideal for automating the initial layout phase of poster creation, reducing manual effort.
- Programmatic Poster Generation: Developers can integrate PosterCopilot-7B into applications requiring dynamic or large-scale poster production.
- Research in Visual AI: Serves as a strong baseline or tool for further research into layout reasoning and controllable graphic design.
For detailed usage and implementation, refer to the code repository and the research paper.